Training a MiniMax-H3 LoRA on a 4090 + 64GB RAM crashes: a systematic diagnosis
marcoc2 · reddit · 2026-08-25
A user training a MiniMax-H3 LoRA with ostris/ai-toolkit on Windows, an RTX 4090 (24GB VRAM) and 64GB RAM hits a native Windows fatal exception: access violation during model loading, with free physical RAM bottoming under 1GB. Their ruled-out list:
- Not a Triton/kernel issue — installed triton-windows, reran 4x, identical crash signature;
- Not classic pagefile OOM — a 0.5s memory watcher shows total commit topped out around 106.5 of 114.4GB; real commit exhaustion throws a clean os error 1455 instead;
- Not offload tuning — layeroffloadingtransformerpercent tested at 1.0/0.5/0.4/0.1, all failing at different points (embedtokens read, VAE init);
- Streaming loads via safeopen were already in place — necessary but not sufficient.
They're asking the community what hardware/config has actually worked for H3 LoRA training.
More from Multimodal
- FLUX 3 Image Klein: 5.6B open-weight image model with native 4K runs on a single consumer GPU — pbaylies · 2026-10-15
- Suno's shareable custom models let collaborators riff, with 2x tuning inputs — suno · 2026-10-15
- Suno's shareable workspaces: team libraries with view/edit access, Pro tier and up — suno · 2026-10-15
- Suno launches collaboration: shareable workspaces and shareable custom models — suno · 2026-10-15
- Kling 4.0 hands-on: native 30-second 1080p generation and Omni Reference tested — HeyZoyaKhan · 2026-10-15
- Kling 4.0 generates a cinematic Tarzan riding a giant wolf, hard to tell it's AI — HeyZoyaKhan · 2026-10-15